Image-to-video is the workflow that separates people who dabble with AI video from people who ship it. Instead of describing a scene in text and hoping the model imagines your product or character correctly, you hand it a still image and ask only for motion. The image locks the look; the prompt controls what happens next. In 2026 every serious model supports image-to-video input, and it has quietly become the default workflow for anyone producing more than one clip.
This guide walks through the full image-to-video pipeline as it works today on veo4.dev: preparing a reference image models actually respect, choosing between reference-image and first-frame inputs, writing shorter motion prompts, and building a consistency system for episodic content. Every capability claim below reflects what each model exposes on our platform at the time of writing. If you have ever generated four versions of "a ceramic mug on a wooden table" and gotten four different mugs, this is the fix.
What Image-to-Video Is and Why It Beats Text-to-Video for Consistency
Text-to-video asks the model to invent everything: subject, lighting, composition, style, and motion. Image-to-video removes most of that invention. You supply a picture, and the model's job shrinks to one task — animate this.
That shrinkage is the consistency advantage. Three cases where image-to-video clearly wins:
Products
A text prompt cannot reliably reproduce your actual product — the exact bottle shape, label typography, specific colorway. An image-to-video generation starting from a clean product photo keeps the real product on screen: for e-commerce and ads, the difference between usable and unusable footage.
Characters
Text-to-video reinvents faces every generation. If you are building a mascot, a virtual presenter, or any recurring character, image-to-video from the same character sheet is the only practical way to keep them recognizable across clips. It is not perfect — faces still drift on longer durations — but it beats re-describing a face in words.
Brand assets
Logos, packaging, a real storefront, a founder's headshot: anything that must look like the real thing has to enter the pipeline as an image. Text descriptions of trademarked visuals produce approximations.
The trade-off is honest: image-to-video constrains the model to what the image contains. When you want a wildly inventive new scene, text-to-video still wins. For everything that needs to look the same twice, start from an image.
Step 1: Prepare a Reference Image the Model Will Respect
Most image-to-video failures are actually input failures. Models follow good images closely and treat bad ones as loose suggestions. Four things matter most.
Resolution and sharpness
Upload the highest-quality source you have — ideally at or above your target output resolution (720p or 1080p on most models here). A soft, compressed, or upscaled image invites the model to hallucinate detail, and hallucinated detail is where products stop looking like your product. Avoid watermarks, text overlays, and JPEG artifacts.
Framing with room to move
A subject cropped tight against all four edges forces the model to either invent surroundings or barely move anything. Frame with breathing room — the composition you want mid-clip, not at its most extreme moment.
Clean, intentional backgrounds
Busy backgrounds multiply the surfaces that can morph. A product on a seamless backdrop, a character against a simple environment, clear depth separation — these animate far more faithfully than clutter. If the background does not matter to the shot, simplify it before generating.
Match the target aspect ratio
Quietly the most important step. Upload a square image and request a 16:9 video, and the model must invent the missing sides — invented regions are where style drift starts. Crop or outpaint your reference to the exact aspect ratio you will generate in: 9:16 for vertical social, 16:9 for landscape. Ten seconds of cropping saves credits on failed generations.
Step 2: Choose an Image-to-Video Model on veo4.dev
Different models accept different kinds of image input, and the distinction matters. A first frame becomes literally the opening frame of your video; a reference image guides the generation more loosely; a last frame defines where the shot ends. Here is what each model exposes on veo4.dev at the time of writing:
| Model | Image input | Required? | Durations | Notes |
|---|---|---|---|---|
| Veo 3.1 | First frame + last frame | Optional | 4 / 6 / 8 sec | The only model here with both endpoints; native audio; 720p/1080p |
| Veo 4 | Reference image | Optional | 4 / 6 / 8 sec | Native audio, negative prompt, seed; 16:9 or 9:16 |
| Kling v2.5 Turbo Pro | First frame | Optional | 5 / 10 sec | Strong motion realism; 16:9, 9:16, or 1:1 |
| Hailuo 2 | First frame | Optional | 6 / 10 sec | Physics-focused; up to 1080p; optional prompt optimizer |
| Seedance 2.0 | Reference image | Optional | 5 / 10 sec | Widest aspect ratio range, including 21:9 and 9:21 |
| Wan AI 2.2 | Reference image | Required | 81–121 frames at adjustable FPS | Dedicated image-to-video model; supports LoRA weights |
| Happy Horse 1.0 | Reference image | Optional | 5 / 10 / 15 sec | Longest single clips here; five aspect ratios |
| Sora 2 | Reference image | Optional | 4 / 8 / 12 sec | Portrait or landscape presets |
How to pick, in practice:
- Controlled start and end points — a transition, a reveal, a loopable shot: Veo 3.1 is the only option with first-plus-last-frame input, and it generates audio natively.
- A guided look with sound: Veo 4 with a reference image, prompt, and optional seed.
- Your image must be the literal opening frame: Kling v2.5 Turbo Pro or Hailuo 2, which take a first frame and preserve it strictly.
- Pure image animation at volume: Wan AI 2.2 is image-to-video only — the reference image is required — and its frame-count, FPS, and LoRA controls make it the tinkerer's choice.
- An unusual aspect ratio: Seedance 2.0 covers everything from 21:9 ultrawide to 9:21 vertical.
You can try all of these from one dashboard on our image-to-video AI page and compare outputs before committing a campaign to one model.
Step 3: Write the Motion Prompt — Shorter Than You Think
The classic mistake when moving to image-to-video: keeping text-to-video habits — long descriptions of subject, lighting, and style the image already carries.
When the image defines the look, your prompt should define only two things:
Motion. What moves, how fast, and where the camera goes. "Slow push-in as steam rises from the cup." "Camera orbits the sneaker 90 degrees, left to right."
Audio, where supported. On Veo 4 and Veo 3.1, which generate native sound, add a short audio direction: "soft cafe ambience, gentle jazz."
That is usually the whole prompt — one or two sentences. Re-describing the image actively hurts: if your text conflicts with the pixels even slightly, you give the model two masters, and the output morphs between what it sees and what it reads. Describe what should change, not what already exists.
A useful template: camera move, then subject action, then environment motion, then audio if supported. Negative prompts (Veo 4, Veo 3.1, Kling) are the right place for "no warping, no distortion, no text" exclusions.
The Series-Consistency Workflow for Episodic and Branded Content
One good clip is luck; ten consistent clips are a workflow. For an episodic series, a product line campaign, or recurring branded content, image-to-video becomes the backbone of a repeatable system:
- Create a canonical asset set. One master image per character, product, or set — same lighting style, same aspect ratio, full resolution. This is your visual bible.
- Generate every clip from the canonical image, never from a previous video frame. Chaining frame-to-frame compounds drift; by clip five your character is a stranger. Always return to the source.
- Keep a prompt bible alongside the image bible. Reuse the same phrasing for recurring camera moves and audio beats.
- Use seeds where available. Veo 4, Veo 3.1, Hailuo 2, Seedance 2.0, Wan AI 2.2, and Happy Horse 1.0 all expose a seed field for near-reproducible re-renders.
- For scene-to-scene continuity, use Veo 3.1's endpoints. Set episode one's closing shot as the first frame of episode two's opener, and the cut becomes seamless.
- Batch-generate, then curate. Generating three takes per shot and picking the best is standard practice, not failure.
Teams comparing platforms should read our best AI image-to-video generator breakdown — the short version: eight models behind one credit system lets you route each shot to whichever model handles it best.
Common Image-to-Video Failures — and Fixes
Three failure patterns account for most wasted credits.
The image gets ignored
The output barely resembles your upload. Usual causes: a low-resolution image, an aspect ratio mismatch forcing major reinvention, or a long prompt that contradicts the image. Fix the input first; shorten the prompt second. On Wan AI 2.2, where the image is the required core input, this failure is rarest.
Morphing away from the reference
The clip starts faithful, then the subject warps — faces shift, labels smear, geometry melts. This worsens with duration and motion intensity. Fixes: request shorter clips (4–6 seconds hold identity far better than 10-plus), calm the motion, use negative prompts against warping, and keep the key subject away from frame edges where models invent freely.
Style drift
The look changes mid-clip — lighting shifts from soft to harsh, a photorealistic product goes slightly illustrated. Common when the prompt implies a style the image does not have. Remove style words entirely and let the image speak. If drift persists, try a different model; styles that break on one often hold on another.
When a generation fails, change one variable at a time — image, prompt, duration, or model — or you will never learn which fix worked.
FAQ
What is the difference between a reference image and a first frame?
A first frame (Veo 3.1, Kling v2.5 Turbo Pro, Hailuo 2) becomes the literal opening frame, so preservation is strict. A reference image (Veo 4, Seedance 2.0, Wan AI 2.2, Happy Horse 1.0, Sora 2) guides subject and look with more interpretive freedom. Use first-frame models when exact pixel continuity matters.
Which models on veo4.dev require an image?
Only Wan AI 2.2 — it is a dedicated image-to-video model, so its reference image field is required. Every other model's image input is optional, so the same model handles both text-to-video and image-to-video jobs.
Can I set both the start and end of a clip?
Yes, on Veo 3.1, which accepts both a first frame and a last frame on our platform — the go-to for transitions, reveals, and stitching consecutive shots. No other model in our lineup currently exposes a last-frame input.
Does image-to-video cost more credits than text-to-video?
No — on veo4.dev credit cost depends on the model, duration, and resolution you choose, not on whether you attach an image. Longer clips and higher resolutions cost more; new users get free starter credits.
What resolution should my reference image be?
At or above your target output — at least 1920x1080 for a 1080p, 16:9 generation. Sharpness matters as much as pixel count: a clean 1080p photo beats a blurry 4K upscale every time.
Can I keep the same character across a whole series?
Mostly, yes — generate every clip from one canonical character image rather than chaining from previous outputs, keep prompts short and consistent, and fix the seed where available. Expect minor variation and curate from multiple takes; perfect consistency is not yet a solved problem.
The Bottom Line
Image-to-video is the highest-leverage habit in AI video right now: it turns an unpredictable creative slot machine into a controllable production tool. Prepare a sharp, well-framed image in the right aspect ratio; pick the input type your shot needs — first frame for strict continuity, reference image for guided looks, first-plus-last on Veo 3.1 for transitions; keep the prompt to motion and audio; and build series on canonical images, not chained outputs.
Models will keep improving, but this image-to-video workflow is durable. The fastest way to internalize it: run the same image through two or three models and watch how differently they move it.